Skill compatibility with providers

Most skills are universal; they work with any provider Cerevisor supports. But some skills are tied to specific providers because they rely on a feature only that provider has.

The skill-preflight gate

Before any workflow runs, Cerevisor's skill-preflight check validates two things for every assigned skill:

1. Existence: the skill is present in your scanned skill folders. If a name doesn't match, Cerevisor suggests the closest matches (deleted or renamed skill folders are the most common cause).
2. Compatibility: the skill declares it works with the resolved provider for each agent that uses it.

If either check fails, the workflow won't start. The error names the specific agent + skill + provider combination that's broken, with a suggested fix (usually: switch the agent's provider, or remove the incompatible skill).

Common incompatibilities

Most skills work everywhere. The ones that don't usually fall into these patterns:

Document and image skills

The document-production skills (docx, pptx, xlsx, pdf) and the image-generation skills (imagegen, gemini-imagegen) are marked in Cerevisor's built-in compatibility table as working only with Claude providers (Anthropic and Claude Code). They depend on a skill runtime (Python scripts + filesystem conventions) that other providers can't follow reliably.

Skills declaring their own restriction

Any skill can declare in its own SKILL.md which providers it supports (see below). Preflight enforces that declaration.

Skills written for Claude Code

Separately from the hard compatibility check, Cerevisor detects when a skill's instructions were written in Claude Code's vocabulary (its tool names, slash commands, subagent references). This detection is advisory, never blocking:

• The chat builder won't automatically attach such a skill to an agent running on Gemini or an OpenAI-compatible provider. When you're driving the session, it asks first and you can attach anyway; workflows you already saved keep running unchanged.
• When such a skill runs on one of Cerevisor's own agents, Cerevisor prepends a short translation note so the agent maps the instructions onto its own tools.

How compatibility is declared

A skill can declare compatibility in its SKILL.md frontmatter with the compatibility: field. Two forms:

---
name: my-skill
description: ...
compatibility: claude-only
---

or an explicit list:

compatibility: [anthropic, openai-compatible]

Recognized presets include universal, claude-only, openai-only, codex-only, and cursor-only. List entries use the provider type names: anthropic, openai-compatible, gemini, codex-cli, cursor-agent, claude-agent-sdk, antigravity-cli.

Resolution order: the skill's own frontmatter wins over Cerevisor's built-in table, which wins over the universal default. That means writing compatibility: universal in a skill's frontmatter overrides a built-in restriction — useful if you've adapted a restricted skill to work elsewhere.

Fixing an incompatibility

When preflight blocks a run, you have three options:

1. Switch the agent's provider

Open the agent's config, change the Provider chip to one that's compatible. Run again.

2. Remove the skill

If the skill isn't critical, remove it from the agent. Run again.

3. Update the skill's compatibility

If you're sure the skill should work with the provider being used (e.g. you're testing a skill against a new provider), edit the skill's SKILL.md frontmatter and set compatibility: accordingly.

Mixing providers within one workflow

This is one of the main reasons Cerevisor lets you set a per-agent provider override. A workflow might have:

• A research agent on Anthropic Sonnet (great at long-form synthesis).
• A writing agent on a local Ollama model (private, free, good enough).
• A publishing agent on Anthropic Haiku (cheap, fast, one-shot task).

The skill-preflight gate runs per-agent, so the workflow runs successfully as long as each agent's assigned skills are compatible with that agent's provider.

When in doubt

Check the skill's compatibility badge in the Skill Workshop ("Universal", "Claude only", "2 of 7", …) — hover it to see exactly which providers the skill works on. If you still can't tell, try it: preflight will catch any compatibility issue cleanly before tokens are spent.